Four ways to automate work.
Anyone looking for AI automation today knows three worlds. The board: You wire up nodes, you are the engineer, and every change means changing the board. The assistant: An agent gets things done for you, live, with your permissions, and rethinks its approach in every session. The company of agents: Roles, an organizational chart, a monthly budget—you are the supervisory board. SubKit is the fourth: AI designs, a person approves, and every run then follows the same path—any authorized person can start it, and it can only be changed through a new version that someone has read.
By category, not by product—so this page does not become outdated with every release from another vendor.
Eight questions, four answers.
| Question | A board for wiring nodes | Personal assistant | Company of agents | SubKit |
|---|---|---|---|---|
| Who builds the workflow? | You, node by node | The model, live during the session | Agents working together | The model designs; a person reads and approves |
| What happens on the second run? | The same thing, as long as you maintain it | A new session that plans again | Agents decide based on the goal and budget | The same path, defined in the approved version; designated AI steps may produce different responses |
| What is it allowed to do? | Everything the connected account is allowed to do | Everything you are allowed to do | Whatever the role and budget allow | Exactly the approved scope for each environment; expanding it requires new approval |
| Can a colleague start it without being able to change it? | Depending on the role | They need their own assistant | Not supported | Yes, from chat, the cockpit or an agent, with their permissions |
| What if a provider changes something? | A red node; you fix it | It improvises | The agent tries again | A proposed fix as a draft; you approve it |
| What did it cost? | Executions counted; AI costs billed by the provider | Included in the subscription or billed by the provider | Monthly budget per agent | Usage per run, workflow and model; the per-tenant budget stops execution at its limit. Infrastructure and providers bill separately |
| Where are data and credentials stored? | With you, if self-hosted | On your computer or in the provider's cloud | With you | With you, encrypted, with test and production separated and support for multiple tenants—calls to connected services and model providers send the data they require to those services |
| Can I prove afterwards what happened? | Execution log | Session history | Audit log | One append-only history across all entry points, recording the person, version and result |
What you still use the others for: The assistant for a one-off task today that needs your computer. The board when you want to draw every branch yourself. The company of agents when autonomy is the goal. SubKit once a task becomes routine and needs to run without you, for others and with a traceable record.
Why not just use a coding agent?
Claude Code, Codex and Hermes are built for developers creating software. Anyone using them to automate a business routine treats it as a programming task: the result is a script that needs to run somewhere—and everything around it is missing. In a team, there is another issue: each person would need their own configured agent.
You would have to build this yourself
- A server, a schedule, containers, updates and someone to check on things at night
- Store credentials securely, build sign-in flows, refresh tokens—none of it in a prompt
- Retry on errors, distinguish between “provider unavailable” and “incorrect data,” and abort instead of writing incorrect data
- Who may start, who may change, who may approve—and a history that documents it
- An interface where a colleague can start it, and a chat where they can ask questions
- A trial run before effects occur, versions, rollback, and test separated from production
- Incoming events with signature verification, waiting for approval, notifications
- Usage per run, and a limit
- And all of it again for the next workflow—a script in a repository is rarely reusable for the next task
Included in SubKit
- A complete stack through the release Compose file—setup, updates and backups remain your operational responsibility
- Connectors with encrypted credentials and sign-in flows
- Consistent retries and error categories for each step
- Roles, scope approvals and append-only history
- Cockpit, chat in Slack, Teams and Google Chat, command line
- Trial run, versions, restore, an environment for each connector
- Signed incoming events, a waiting step with a deadline, notification rules
- Usage per run, budget per tenant
- Every workflow, connector and skill is reusable—with a description and category that your agent can also read
Both worlds together: Your coding agent remains your tool for software. Your company's routine operations are not its proper job—which is why SubKit is a tool it can use: it designs and starts workflows there, while SubKit handles permissions, approvals and history. Shell access, using the file system as a workspace, unrestricted long-running jobs and bulk data migrations remain the domain of developer tools.
See the difference for yourself.
What your team can set up with it
Specific situations and the workflows you set up for them.
Use cases →